Manuscript_fishing_hours

Open R libraries

library(tidyverse)
library(tidyr)
library(ggplot2)
library(data.table)
library(gridExtra)
library(maps)
library(raster)
library(sf)
library(viridis)
library(scales)
library(dplyr)
library(randomForest)
library(caret)
library(pdp)
library(knitr)
library(kableExtra)
library(future)
library(spdep)
library(ncf)
library(blockCV)
library(parallel)
library(doParallel)
library(sp)
library(purrr)
library(kableExtra)
library(webshot2)

Open datasets

#AIS data
load(here::here("R","Data","AIS_fishing.Rdata"))

#SAR data
load(here::here("R","Data","SAR_fishing.Rdata"))

# Open the saved adjusted rasters
Ports_adjusted <- raster(here::here("R", "Data", "distance-from-port-0.1deg-adjusted.tif"))
Shore_adjusted <- raster(here::here("R", "Data", "distance-from-shore-0.1deg-adjusted.tif"))
Bathy_adjusted <- raster(here::here("R", "Data", "bathymetry-0.1deg-adjusted.tif"))

#
load(here::here("R", "Data","combined_data_O1deg.Rdata"))

# Load the saved RF models
rf_model_no_transform <- readRDS(here::here("R", "Data", "rf_model_no_transform.rds"))
rf_model_original <- readRDS(here::here("R", "Data", "rf_model_original.rds"))
rf_model_log <- readRDS(here::here("R", "Data", "rf_model_log.rds"))
rf_model_fishing_log <- readRDS(here::here("R", "Data", "rf_model_fishing_log.rds"))

AIS data

Data from Kroodsma et al. (2018) Science, accessible at: Global Fishing Watch Data Download Portal.

The data is accessible up to the end of the year 2020, we used four entire years of data (2017, 2018, 2019, 2020) to match SAR data records.

# Set the path to the 2017-2020 folder

#path <- "R/Data/AIS Fishing Effort 2017-2020"

# List all CSV files in the folder
#AIS_csv_files <- list.files(path = here::here(path), pattern = "*.csv", full.names = TRUE, recursive = TRUE)

# Read all CSV files and combine them into a single data frame
#AIS_fishing <- AIS_csv_files %>%
#  map_df(~read_csv(.))

# Aggregate fishing hours by latitude and longitude
aggregated_AIS_fishing <- AIS_fishing %>%
  group_by(cell_ll_lat, cell_ll_lon) %>%
  summarise(total_fishing_hours = sum(fishing_hours, na.rm = TRUE)) %>%
  ungroup() %>%
  # Remove any cells with zero or negative fishing hours
  filter(total_fishing_hours > 0)

# Function to standardize coordinates to 0.1 degree resolution
standardize_coords <- function(lon, lat) {
  list(
    lon_std = floor(lon * 10) / 10,
    lat_std = floor(lat * 10) / 10
  )
}

# Standardize and aggregate AIS data
AIS_data_std <- aggregated_AIS_fishing %>%
  mutate(coords = map2(cell_ll_lon, cell_ll_lat, standardize_coords)) %>%
  mutate(
    lon_std = map_dbl(coords, ~ .x$lon_std),
    lat_std = map_dbl(coords, ~ .x$lat_std)
  ) %>%
  group_by(lon_std, lat_std) %>%
  summarise(total_fishing_hours = sum(total_fishing_hours, na.rm = TRUE), .groups = "drop")

# Create the world map
world_map <- map_data("world")

# Assign plot to variable
AIS_data_plot <- ggplot() +
  geom_map(data = world_map, map = world_map,
           aes(long = long, lat = lat, map_id = region),
           color = "black", fill = "lightgray", size = 0.1) +
  geom_tile(data = AIS_data_std, 
            aes(x = lon_std, y = lat_std, fill = total_fishing_hours)) +
  scale_fill_viridis(
    option = "inferno",
    direction = -1,
    trans = "log1p",
    name = "AIS fishing effort (hours; 2017-2020)", 
    breaks = c(0, 1, 10, 100, 1000, 10000, 100000),
    labels = scales::comma,
    guide = guide_colorbar(barwidth = 20, barheight = 0.5, title.position = "top", title.hjust = 0.5)
  ) +
  coord_fixed(1.3) +
  theme_minimal() +
  labs(title = NULL,
       x = "Longitude", y = "Latitude") +
  theme(
    legend.position = "bottom",
    legend.direction = "horizontal",
    legend.box = "vertical",
    legend.margin = ggplot2::margin(t = 20, r = 0, b = 0, l = 0),
    legend.title = element_text(margin = ggplot2::margin(b = 10))
  )

print(AIS_data_plot)

# Save the plot
ggsave(here::here("R","Outputs","AIS_fishing_map.png"), 
       plot = AIS_data_plot,
       width = 12, 
       height = 8, 
       dpi = 300)

Sentinel-1 SAR data

Data from Paolo et al. 2024 Nature, accessible at: Global Fishing Watch SAR Vessel Detections

The data is accessible from 2017 to today, we used four entire years of data (2017, 2018, 2019, 2020) to match AIS records.

# Set the path to the 2016 folder
#path <- "R/Data/SAR Vessel detections 2017-2020"

# List all CSV files in the folder
#SAR_csv_files <- list.files(path = here::here(path), pattern = "*.csv", full.names = TRUE)

# Read all CSV files and combine them into a single data frame
#SAR_fishing <- SAR_csv_files %>%
#  map_df(~read_csv(.))

# Aggregate fishing hours by latitude and longitude
aggregated_SAR_fishing <- SAR_fishing %>%
  mutate(
    lat_rounded = round(lat, digits = 2),
    lon_rounded = round(lon, digits = 2)
  ) %>%
  group_by(lat_rounded, lon_rounded) %>%
  filter(fishing_score >= 0.9) %>%
  summarise(
    total_presence_score = sum(presence_score, na.rm = TRUE),
    avg_fishing_score = mean(fishing_score, na.rm = TRUE),
    count = n()
  ) %>%
  mutate(total_presence_score = round(total_presence_score, digits = 0)) %>%
  ungroup()

# Standardize and aggregate SAR data
SAR_data_std <- aggregated_SAR_fishing %>%
  mutate(coords = map2(lon_rounded, lat_rounded, standardize_coords)) %>%
  mutate(
    lon_std = map_dbl(coords, ~ .x$lon_std),
    lat_std = map_dbl(coords, ~ .x$lat_std)
  ) %>%
  group_by(lon_std, lat_std) %>%
  summarise(total_presence_score = sum(total_presence_score, na.rm = TRUE), .groups = "drop")

# Create the world map
world_map <- map_data("world")

# Assign plot to variable
SAR_data_plot <- ggplot() +
  geom_map(data = world_map, map = world_map,
           aes(long = long, lat = lat, map_id = region),
           color = "black", fill = "lightgray", size = 0.1) +
  geom_tile(data = SAR_data_std, 
            aes(x = lon_std, y = lat_std, fill = total_presence_score)) +
  scale_fill_viridis(
    option = "inferno",
    direction = -1,
    trans = "log1p",
    name = "Fishing vessel detections (2017-2020)", 
    breaks = c(0, 1, 10, 100, 1000, 10000),
    labels = scales::comma,
    guide = guide_colorbar(barwidth = 20, barheight = 0.5, title.position = "top", title.hjust = 0.5)
  ) +
  coord_fixed(1.3) +
  theme_minimal() +
  labs(title = NULL,
       x = "Longitude", y = "Latitude") +
  theme(
    legend.position = "bottom",
    legend.direction = "horizontal",
    legend.box = "vertical",
    legend.margin = ggplot2::margin(t = 20, r = 0, b = 0, l = 0),
    legend.title = element_text(margin = ggplot2::margin(b = 10))
  )
print(SAR_data_plot)

# Save the plot
ggsave(here::here("R","Outputs", "SAR_fishing_map.png"), 
       plot = SAR_data_plot,
       width = 12, 
       height = 8, 
       dpi = 300)

Compare AIS, and SAR detection locations

Identifying grid cells with only AIS, only SAR detections or both data types.

# Merge the datasets
combined_data <- full_join(
  AIS_data_std,
  SAR_data_std,
  by = c("lon_std", "lat_std")
)

# Categorize each cell
combined_data <- combined_data %>%
  mutate(category = case_when(
    total_fishing_hours > 0 & total_presence_score > 0 ~ "Both AIS and SAR",
    total_fishing_hours > 0 & (is.na(total_presence_score) | total_presence_score == 0) ~ "Only AIS",
    (is.na(total_fishing_hours) | total_fishing_hours == 0) & total_presence_score > 0 ~ "Only SAR",
    TRUE ~ "No fishing detected"
  ))

# Create the world map
world_map <- map_data("world")

# Create the plot
world_plot <- ggplot() +
  geom_map(data = world_map, map = world_map,
           aes(long = long, lat = lat, map_id = region),
           color = "black", fill = "lightgray", size = 0.1) +
  geom_tile(data = combined_data, 
            aes(x = lon_std, y = lat_std, fill = category)) +
  scale_fill_manual(
    values = c("Both AIS and SAR" = "purple", 
               "Only AIS" = "blue", 
               "Only SAR" = "red", 
               "No fishing detected" = "white"),
    name = "Fishing data source",
    guide = guide_legend(title.position = "top", title.hjust = 0.5)
  ) +
  coord_fixed(1.3) +
  theme_minimal() +
  labs(title = "Global fishing detection",
       subtitle = "Comparison of AIS (2017-2020) and SAR (2017-2020) data at 0.1-degree resolution",
       x = "Longitude", y = "Latitude") +
  theme(
    legend.position = "bottom",
    legend.direction = "horizontal",
    legend.box = "vertical",
    legend.margin = ggplot2::margin(t = 20, r = 0, b = 0, l = 0),
    legend.title = element_text(margin = ggplot2::margin(b = 10))
  )

# Display the plot
print(world_plot)

# Save the plot
ggsave(here::here("R","Outputs", "AIS_SAR_comparison_map.png"), 
       plot = world_plot,
       width = 12, 
       height = 8, 
       dpi = 300)

# Get summary statistics
summary_stats <- combined_data %>% 
  count(category) %>% 
  mutate(percentage = n / sum(n) * 100) %>%
  rename(`Number of cells` = n) %>%
  mutate(percentage = round(percentage, 2))

# Create the table
table_output <- kable(summary_stats, 
      format = "html", 
      col.names = c("Category", "Number of cells", "Percentage (%)"),
      caption = "Summary statistics of data categories") %>%
  kable_styling(bootstrap_options = c("striped", "hover", "condensed", "responsive"),
                full_width = FALSE, 
                position = "center") %>%
  column_spec(2, width = "150px") %>%
  column_spec(3, width = "150px")

table_output

Summary statistics of data categories

Category	Number of cells	Percentage (%)
Both AIS and SAR	163095	9.12
Only AIS	1566190	87.60
Only SAR	58668	3.28

# Save as PNG
save_kable(table_output, file = here::here("R","Outputs","summary_stats_table.png"))

Random forest model

Predicting fishing hours in areas with only SAR data at a 0.1 degree resolution. Using a random forest model of >160,000 associations between SAR vessel detections and AIS fishing hours globally, geographical coordinates, distance to ports, distance to shore and bathymetry.

# Stack the resampled rasters
raster_stack <- stack(Shore_adjusted, Ports_adjusted, Bathy_adjusted)

# Convert the stack to a dataframe
raster_df <- as.data.frame(raster_stack, xy = TRUE)

# Rename the columns
names(raster_df) <- c("x", "y", "dist_shore", "dist_ports", "bathy")

# Remove NA values if desired
raster_df <- na.omit(raster_df)

# Convert to data.table for efficiency
setDT(raster_df)

# Round x and y to 1 decimal place for consistency
raster_df[, `:=`(
  lon_std = round(x, digits = 1),
  lat_std = round(y, digits = 1)
)]

# Keep only ocean areas (negative bathymetry values)
raster_df <- raster_df[bathy < 0]

# Keep the first occurrence of each coordinate pair (because there are duplicates)
raster_df <- unique(raster_df, by = c("lon_std", "lat_std"))
setDT(raster_df)

# Now proceed with the join and model training
#load(here::here("R,"Data","combined_data_O1deg.Rdata"))

# Keep the first occurrence of each coordinate pair (because there are duplicates)
combined_data_01deg <- unique(combined_data_01deg, by = c("lon_std", "lat_std"))
setDT(combined_data_01deg)

# Perform the join using data.table
combined_data_with_rasters <- raster_df[combined_data_01deg, 
                                        on = .(lon_std, lat_std), 
                                        nomatch = 0]

# Convert back to dataframe if needed
combined_data_with_rasters <- as.data.frame(combined_data_with_rasters)

# Create the new combined dataframe
combined_data_all <- raster_df[, .(lon_std, lat_std, dist_shore, dist_ports, bathy)]

# Perform the join
combined_data_all <- merge(combined_data_all, combined_data_01deg, 
                           by = c("lon_std", "lat_std"), 
                           all.x = TRUE)

# Fill NA values for has_AIS and has_SAR with FALSE
combined_data_all[is.na(has_AIS), has_AIS := FALSE]
combined_data_all[is.na(has_SAR), has_SAR := FALSE]

# Categorize each cell
combined_data_all <- combined_data_all %>%
  mutate(category = case_when(
    has_AIS == TRUE ~ "AIS Data",
    has_SAR == TRUE ~ "SAR Data Only",
    TRUE ~ "No AIS or SAR Data"
  ))

# Create the world map
world_map <- map_data("world")

# Create the plot
world_plot <- ggplot() +
  geom_map(data = world_map, map = world_map,
           aes(long = long, lat = lat, map_id = region),
           color = "black", fill = "lightgray", size = 0.1) +
  geom_tile(data = combined_data_all, 
            aes(x = lon_std, y = lat_std, fill = category, color = category)) +
  scale_fill_manual(
    values = c("AIS Data" = "blue", 
               "SAR Data Only" = "red", 
               "No AIS or SAR Data" = "black"),
    name = "Data Source",
    guide = guide_legend(title.position = "top", title.hjust = 0.5)
  ) +
  scale_color_manual(
    values = c("AIS Data" = "blue", 
               "SAR Data Only" = "red", 
               "No AIS or SAR Data" = "black"),
    guide = "none"
  ) +
  coord_fixed(1.3) +
  theme_minimal() +
  labs(title = "Global Data Coverage",
       subtitle = "Distribution of AIS and SAR data at 0.1-degree resolution",
       x = "Longitude", y = "Latitude") +
 theme(
    legend.position = "bottom",
    legend.direction = "horizontal",
    legend.box = "vertical",
    legend.margin = ggplot2::margin(t = 20, r = 0, b = 0, l = 0),
    legend.title = element_text(margin = ggplot2::margin(b = 10)),
    panel.grid.major = element_blank(),
    panel.grid.minor = element_blank(),
    axis.text = element_text(size = 8),
    axis.title = element_text(size = 10),
    plot.title = element_text(size = 14, face = "bold"),
    plot.subtitle = element_text(size = 12)
  )

# Print the plot
#print(world_plot)

# Calculate summary statistics
summary_stats <- combined_data_all[, .(
  data_type = case_when(
    has_AIS == TRUE ~ "AIS Data",
    has_SAR == TRUE ~ "SAR Data Only",
    TRUE ~ "No AIS or SAR Data"
  )
)][, .(
  num_cells = .N,
  percentage = .N / nrow(combined_data_all) * 100
), by = data_type]

# Order the data types
summary_stats <- summary_stats[order(match(data_type, c("AIS Data", "SAR Data Only", "No AIS or SAR Data")))]

# Add total row
total_row <- data.table(
  data_type = "Total",
  num_cells = sum(summary_stats$num_cells),
  percentage = 100
)

summary_stats <- rbindlist(list(summary_stats, total_row))

# Create kable
kable_output <- summary_stats %>%
  kable(
    col.names = c("Data Type", "Number of Cells", "Percentage (%)"),
    digits = c(0, 0, 2),
    align = c("l", "r", "r"),
    caption = "Summary Statistics of Data Types"
  ) %>%
  kable_styling(bootstrap_options = c("striped", "hover", "condensed"), full_width = FALSE) %>%
  row_spec(nrow(summary_stats), bold = TRUE, background = "#F0F0F0") %>%
  footnote(
    general = "AIS Data category includes cells with both AIS and SAR data.",
    general_title = "Note:",
    footnote_as_chunk = TRUE
  )

# Print the kable
#kable_output

# Prepare the training data
training_data <- combined_data_with_rasters %>%
  filter(has_AIS & has_SAR) %>%
  dplyr::select(total_fishing_hours, total_presence_score, lon_std, lat_std, dist_shore, dist_ports, bathy) %>%
  na.omit()

Comparison of transformations in models

# Prepare the data
#load(here::here("R","Data","training_data.Rdata"))

#training_data_log <- training_data %>%
#  mutate(
#    log_total_presence_score = log10(total_presence_score + 1),
#    log_total_fishing_hours = log10(total_fishing_hours + 1)
#  )

# Function to run a single model
#run_model <- function(formula, data) {
#  randomForest(
#    formula,
#    data = data,
#    ntree = 500,
#    importance = TRUE
#  )
#}

# Set up parallel processing
#num_cores <- detectCores() - 1  # Use all but one core
#cl <- makeCluster(num_cores)

# Export necessary objects to the cluster
#clusterExport(cl, c("training_data_log", "run_model"))

# Load required packages on each cluster
#clusterEvalQ(cl, library(randomForest))

# Define the models
#models <- list(
#  no_transform = as.formula(total_fishing_hours ~ total_presence_score + lon_std + lat_std + dist_shore + #dist_ports + bathy),
#  original = as.formula(total_fishing_hours ~ log_total_presence_score + lon_std + lat_std + dist_shore + #dist_ports + bathy),
#  log = as.formula(log_total_fishing_hours ~ log_total_presence_score + lon_std + lat_std + dist_shore + #dist_ports + bathy)
#)

# Run models in parallel
#results <- parLapply(cl, models, function(formula) run_model(formula, training_data_log))

# Stop the cluster
#stopCluster(cl)

# Save the models
#rf_model_no_transform <- results[[1]]
#rf_model_original <- results[[2]]
#rf_model_log <- results[[3]]

# Save models to files
#saveRDS(rf_model_no_transform, "rf_model_no_transform.rds")
#saveRDS(rf_model_original, "rf_model_original.rds")
#saveRDS(rf_model_log, "rf_model_log.rds")

# Add the new model with only total_fishing_hours log-transformed
#rf_model_fishing_log <- randomForest(
#  log_total_fishing_hours ~ total_presence_score + lon_std + lat_std + dist_shore + dist_ports + bathy,
#  data = training_data_log,
#  ntree = 500,
#  importance = TRUE
#)

# Save the new model
#saveRDS(rf_model_fishing_log, "rf_model_fishing_log.rds")

# Function to evaluate models
evaluate_model <- function(model, data, log_target = FALSE) {
  predictions <- predict(model, newdata = data)
  if (log_target) {
    predictions <- 10^predictions - 1
  }
  
  actual <- data$total_fishing_hours
  
  # Basic Error Metrics
  mae <- mean(abs(actual - predictions), na.rm = TRUE)
  rmse <- sqrt(mean((actual - predictions)^2, na.rm = TRUE))
  mape <- mean(abs((actual - predictions) / actual) * 100, na.rm = TRUE)
  medae <- median(abs(actual - predictions), na.rm = TRUE)
  
  # R-squared (matching randomForest's % Var explained)
  r_squared <- model$rsq[length(model$rsq)]
  
  # Adjusted R-squared
  n <- length(actual)
  p <- length(model$forest$independent.variable.names) # Number of predictors
  adj_r_squared <- 1 - ((1 - r_squared) * (n - 1) / (n - p - 1))
  
  # Residual Analysis
  residuals <- actual - predictions
  mean_residual <- mean(residuals, na.rm = TRUE)
  sd_residual <- sd(residuals, na.rm = TRUE)
  
  # Feature Importance (for Random Forest)
  feature_importance <- importance(model)
  
  return(list(
    "Mean Absolute Error" = mae,
    "Root Mean Squared Error" = rmse,
    "Mean Absolute Percentage Error" = mape,
    "Median Absolute Error" = medae,
    "R-Squared" = r_squared,
    "Adjusted R-Squared" = adj_r_squared,
    "Mean of Residuals" = mean_residual,
    "Standard Deviation of Residuals" = sd_residual,
    "Feature Importance" = feature_importance
  ))
}

# Evaluate all models
validation_data <- combined_data_with_rasters %>%
  mutate(
    data_category = case_when(
      has_AIS & has_SAR ~ "Both AIS and SAR",
      has_AIS & !has_SAR ~ "Only AIS",
      !has_AIS & has_SAR ~ "Only SAR",
      TRUE ~ "No fishing detected"
    )
  )

validation_data <- validation_data %>% filter(data_category == "Both AIS and SAR")

# Evaluate all models
results_no_transform <- evaluate_model(rf_model_no_transform, validation_data)

validation_data_logpres <- validation_data %>%
  mutate(log_total_presence_score = log10(total_presence_score + 1))

results_original <- evaluate_model(rf_model_original, validation_data_logpres)

# Add evaluation for the new model (fishing hours log-transformed)
results_fishing_log <- evaluate_model(rf_model_fishing_log, validation_data, log_target = TRUE)

results_log <- evaluate_model(rf_model_log, validation_data_logpres, log_target = TRUE)

# Create a data frame with the results
results_df <- data.frame(
  Metric = c(
  "Mean absolute error",                    # MAE
  "Root mean squared error",                # RMSE
  "Mean absolute percentage error",         # MAPE
  "Median absolute error",                  # MdAE
  "R-squared",                              # R² or R2
  "Adjusted R-squared",                     # Adj. R²
  "Mean of residuals",
  "Standard deviation of residuals"
),
  No_Transform = c(results_no_transform$`Mean Absolute Error`, results_no_transform$`Root Mean Squared Error`, results_no_transform$`Mean Absolute Percentage Error`, results_no_transform$`Median Absolute Error`, results_no_transform$`R-Squared`, results_no_transform$`Adjusted R-Squared`, results_no_transform$`Mean of Residuals`, results_no_transform$`Standard Deviation of Residuals`),
  Fishing_Log = c(results_fishing_log$`Mean Absolute Error`, results_fishing_log$`Root Mean Squared Error`, results_fishing_log$`Mean Absolute Percentage Error`, results_fishing_log$`Median Absolute Error`, results_fishing_log$`R-Squared`, results_fishing_log$`Adjusted R-Squared`, results_fishing_log$`Mean of Residuals`, results_fishing_log$`Standard Deviation of Residuals`),
  Presence_Log = c(results_original$`Mean Absolute Error`, results_original$`Root Mean Squared Error`, results_original$`Mean Absolute Percentage Error`, results_original$`Median Absolute Error`, results_original$`R-Squared`, results_original$`Adjusted R-Squared`, results_original$`Mean of Residuals`, results_original$`Standard Deviation of Residuals`),
  Both_Log = c(results_log$`Mean Absolute Error`, results_log$`Root Mean Squared Error`, results_log$`Mean Absolute Percentage Error`, results_log$`Median Absolute Error`, results_log$`R-Squared`, results_log$`Adjusted R-Squared`, results_log$`Mean of Residuals`, results_log$`Standard Deviation of Residuals`)
)

# Create and save the table as HTML
table_output <- kable(results_df, format = "html", digits = 3,
      col.names = c("Metric", "No transform", "Fishing hours log", "Presence score log", "Both log"),
      caption = "Model performance comparison") %>%
  kable_styling(bootstrap_options = c("striped", "hover", "condensed", "responsive"),
                full_width = FALSE) %>%
  add_header_above(c(" " = 1, "Models" = 4)) %>%
  column_spec(1, bold = TRUE)

# Save as HTML first
save_kable(table_output, file = here::here("R", "Outputs", "model_performance.html"))

# Then screenshot with custom viewport
webshot(here::here("R", "Outputs", "model_performance.html"), 
        here::here("R", "Outputs", "model_performance.png"),
        vwidth = 800,   # Width in pixels
        vheight = 400,  # Height in pixels - adjust as needed
        zoom = 2)

Interpretation of model comparison metrics

Based on the provided performance metrics, I would choose the Fishing Hours Log-Transformed Model. Here’s the reasoning:

1. R-Squared and Adjusted R-Squared: The Fishing Hours Log model has the highest R-squared (0.8239) and Adjusted R-squared values, indicating it explains the most variance in the data.
2. Mean Absolute Percentage Error (MAPE): This model has a significantly lower MAPE (69.71%) compared to the No Transform and Presence Log models (both over 1200%). This suggests much better relative accuracy. Median Absolute Error: It has the lowest median absolute error (10.19), which indicates good performance on typical cases.
3. Root Mean Squared Error (RMSE): While higher than the No Transform model, the difference isn’t as dramatic as the improvement in MAPE.
4. Mean Absolute Error (MAE): Although higher than No Transform and Presence Log models, this should be considered in context with other metrics.

The Both Log model performs very similarly to the Fishing Hours Log model, but the latter edges it out slightly in most metrics.

The No Transform and Presence Log models, despite having lower MAE and RMSE, have extremely high MAPE values, suggesting they might be making large relative errors, especially on smaller values.

The logarithmic transformation of fishing hours seems to have addressed some issues with the data distribution, leading to more balanced performance across different scales of the target variable.

In conclusion, the Fishing Hours Log-Transformed Model appears to offer the best overall performance, particularly in terms of explained variance and relative error metrics. However, the choice might depend on the specific requirements of your application, such as whether absolute or relative errors are more important in your context.

Selected Model performance

evaluate_model <- function(model, data, log_target = FALSE) {
  predictions <- predict(model, newdata = data)
  if (log_target) {
    predictions <- 10^predictions - 1
  }
  
  actual <- if (log_target) 10^data$log_total_fishing_hours - 1 else data$total_fishing_hours
  
  # Basic Error Metrics
  mae <- mean(abs(actual - predictions), na.rm = TRUE)
  rmse <- sqrt(mean((actual - predictions)^2, na.rm = TRUE))
  mape <- mean(abs((actual - predictions) / actual) * 100, na.rm = TRUE)
  medae <- median(abs(actual - predictions), na.rm = TRUE)
  
  # R-squared (matching randomForest's % Var explained)
  r_squared <- model$rsq[length(model$rsq)]
  
  # Adjusted R-squared
  n <- length(actual)
  p <- length(model$forest$independent.variable.names)
  adj_r_squared <- 1 - ((1 - r_squared) * (n - 1) / (n - p - 1))
  
  # Residual Analysis
  residuals <- actual - predictions
  mean_residual <- mean(residuals, na.rm = TRUE)
  sd_residual <- sd(residuals, na.rm = TRUE)
  
  # Feature Importance (for Random Forest)
  feature_importance <- importance(model)
  
  return(list(
    "Mean absolute error" = mae,
    "Root mean squared error" = rmse,
    "Mean absolute percentage error" = mape,
    "Median absolute error" = medae,  # FIXED TYPO
    "R-squared" = r_squared,
    "Adjusted R-squared" = adj_r_squared,  # FIXED CASE
    "Mean of residuals" = mean_residual,
    "Standard deviation of residuals" = sd_residual,
    "Feature importance" = feature_importance
  ))
}

validation_data <- combined_data_with_rasters %>%
  mutate(
    data_category = case_when(
      has_AIS & has_SAR ~ "Both AIS and SAR",
      has_AIS & !has_SAR ~ "Only AIS",
      !has_AIS & has_SAR ~ "Only SAR",
      TRUE ~ "No fishing detected"
    ),
    log_total_fishing_hours = log10(total_fishing_hours + 1)
  )

# Evaluate the model
validation_data <- validation_data %>% filter(data_category == "Both AIS and SAR")
results_rf_fishing_log <- evaluate_model(rf_model_fishing_log, validation_data, log_target = TRUE)

# Create a dataframe for the table
results_table <- data.frame(
  Metric = c("Mean absolute error", "Root mean squared error", "Mean absolute percentage error",
             "Median absolute error", "R-squared", "Adjusted R-squared",
             "Mean of residuals", "Standard deviation of residuals"),
  Value = round(c(results_rf_fishing_log$`Mean absolute error`,
                  results_rf_fishing_log$`Root mean squared error`,
                  results_rf_fishing_log$`Mean absolute percentage error`,
                  results_rf_fishing_log$`Median absolute error`,  # FIXED: was "Median ebsolute error"
                  results_rf_fishing_log$`R-squared`,
                  results_rf_fishing_log$`Adjusted R-squared`,  # FIXED: was "Adjusted r-squared"
                  results_rf_fishing_log$`Mean of residuals`,
                  results_rf_fishing_log$`Standard deviation of residuals`),
                2))

table_output <- kable(results_table, format = "html", digits = 4, 
      caption = "Model evaluation metrics for log-transformed fishing hours model") %>%
  kable_styling(bootstrap_options = c("striped", "hover", "condensed", "responsive"),
                full_width = TRUE,
                position = "center")

# Save as HTML first
save_kable(table_output, file = here::here("R", "Outputs", "model_evaluation.html"))

# Then screenshot with custom viewport
webshot(here::here("R", "Outputs", "model_evaluation.html"), 
        here::here("R", "Outputs", "model_evaluation.png"),
        vwidth = 800,   # Width in pixels
        vheight = 360,  # Height in pixels - adjust as needed
        zoom = 2)

# For feature importance, create a separate table
# Use $ before the backticks
feature_importance <- as.data.frame(results_rf_fishing_log$`Feature importance`)
feature_importance$Feature <- rownames(feature_importance)
feature_importance <- feature_importance[, c("Feature", "%IncMSE", "IncNodePurity")]
colnames(feature_importance) <- c("Feature", "%IncMSE", "IncNodePurity")

# Sort the feature importance table by %IncMSE in descending order
feature_importance <- feature_importance[order(-feature_importance$`%IncMSE`), ]

# Create the table
table_output <- kable(feature_importance, format = "html", digits = 4, 
      col.names = c("Feature", "%IncMSE", "IncNodePurity"),
      caption = "Feature importance for log-transformed fishing hours model") %>%
  kable_styling(bootstrap_options = c("striped", "hover", "condensed", "responsive"),
                full_width = TRUE, position = "center") %>%
  column_spec(1, bold = TRUE) %>%
  column_spec(2:3, width = "150px")

# Save as HTML first
save_kable(table_output, file = here::here("R", "Outputs", "feature_importance.html"))

# Then screenshot with custom viewport
webshot(here::here("R", "Outputs", "feature_importance.html"), 
        here::here("R", "Outputs", "feature_importance.png"),
        vwidth = 720,   # Width in pixels
        vheight = 300,  # Height in pixels - adjust as needed
        zoom = 2)

Maps of predictions

# Prepare the prediction data
prediction_data <- combined_data_with_rasters %>%
  dplyr::select(total_presence_score, lon_std, lat_std, dist_shore, dist_ports, bathy)

# Make predictions
log_predictions <- predict(rf_model_fishing_log, newdata = prediction_data)

# Back-transform predictions
predictions <- 10^log_predictions - 1

# Add predictions to the original dataset
combined_data_01deg <- combined_data_01deg %>%
  mutate(
    predicted_fishing_hours = case_when(
      has_AIS ~ total_fishing_hours,
      has_SAR ~ predictions[match(paste(lon_std, lat_std), paste(prediction_data$lon_std, prediction_data$lat_std))],
      TRUE ~ 0
    )
  )

#Violin plot of observed versus predicted fishing hours 
# Prepare data for plotting
plot_data <- combined_data_01deg %>%
  mutate(
    ais_fishing_hours = if_else(has_AIS, total_fishing_hours, NA_real_),
    sar_predicted_hours = if_else(!has_AIS & has_SAR, predicted_fishing_hours, NA_real_)
  ) %>%
  dplyr::select(ais_fishing_hours, sar_predicted_hours) %>%
  pivot_longer(cols = c(ais_fishing_hours, sar_predicted_hours),
               names_to = "type",
               values_to = "hours") %>%
  filter(!is.na(hours))

# Create violin plot
violin_plot <- ggplot(plot_data, aes(x = type, y = hours, fill = type)) +
  geom_violin(trim = FALSE) +
  geom_boxplot(width = 0.1, fill = "white", color = "black", alpha = 0.5, outlier.shape = NA) +
  scale_y_log10(labels = scales::comma_format(accuracy = 1)) +
  scale_x_discrete(labels = c("ais_fishing_hours" = "AIS Data", 
                              "sar_predicted_hours" = "SAR Predictions")) +
  labs(title = "Comparison of AIS Fishing Hours and SAR Predicted Fishing Hours",
       subtitle = "AIS data for AIS-covered areas, Predictions for SAR-only areas",
       x = "",
       y = "Fishing Hours (log scale)",
       fill = "Type") +
  theme_minimal() +
  theme(legend.position = "none",
        axis.text.x = element_text(angle = 45, hjust = 1))

# Print the plot
print(violin_plot)

# Map of predicted fishing hours only 
# Prepare the data for the map
map_data <- combined_data_01deg %>%
  filter(!has_AIS & has_SAR) %>%
  dplyr::select(lon_std, lat_std, predicted_fishing_hours)

# The map_data now contains back-transformed predictions, so no further transformation is needed

#predicted_SAR_only_1RF=map_data
#save(predicted_SAR_only_1RF, file="predicted_SAR_only_1RF.Rdata")

# Create the world map
world_map <- map_data("world")

# Function to create map for a specific region
create_region_map <- function(data, world_map, lon_col, lat_col, lon_range, lat_range, title, subtitle) {
  ggplot() +
    geom_map(data = world_map, map = world_map,
             aes(long, lat, map_id = region),
             color = "darkgray", fill = "lightgray", size = 0.1) +
    geom_tile(data = data, 
              aes(x = .data[[lon_col]], y = .data[[lat_col]], fill = predicted_fishing_hours)) +
    scale_fill_viridis(
      option = "inferno",
      direction = -1,
      trans = "log1p",
      name = "Predicted fishing hours (2017-2020)", 
      breaks = c(0, 1, 10, 100, 1000, 10000, 100000, 1000000),
      labels = scales::comma,
      guide = guide_colorbar(barwidth = 20, barheight = 0.5, title.position = "top", title.hjust = 0.5)
    ) +
    coord_fixed(1.3, xlim = lon_range, ylim = lat_range) +
    theme_minimal() +
    labs(title = title,
         subtitle = subtitle,
         x = "Longitude", y = "Latitude") +
    theme(
      legend.position = "bottom",
      legend.direction = "horizontal",
      legend.box = "vertical",
      legend.margin = ggplot2::margin(t = 20, r = 0, b = 0, l = 0),
      legend.title = element_text(margin = ggplot2::margin(b = 10))
    )
}

# Global map
predicted_SAR_only_plot <- create_region_map(map_data, world_map, "lon_std", "lat_std", c(-180, 180), c(-90, 90), "Predicted Fishing Hours in Areas with Only SAR Detections", "0.1 degree resolution")

# Caribbean map
caribbean_map <- create_region_map(map_data, world_map, "lon_std", "lat_std", c(-100, -50), c(0, 40), "Predicted Fishing Hours in the Caribbean", "0.1 degree resolution")

# Northwestern Indian Ocean to Western European waters map
indian_european_map <- create_region_map(map_data, world_map, "lon_std", "lat_std", c(-20, 80), c(0, 70), "Predicted Fishing Hours from Northern Indian Ocean \nto Eastern Atlantic", "0.1 degree resolution")

# Asia map
asia_map <- create_region_map(map_data, world_map, "lon_std", "lat_std", c(60, 180), c(-20, 60), "Predicted Fishing Hours in Asia", "0.1 degree resolution")

# Print the maps
#print(predicted_SAR_only_plot)
print(caribbean_map)

print(indian_european_map)

print(asia_map)

#Map of both original and predicted AIS fishing hours 
# Visualize the results
predicted_plot <- ggplot() +
  geom_map(data = world_map, map = world_map,
           aes(long, lat, map_id = region),
           color = "black", fill = "lightgray", size = 0.1) +
  geom_tile(data = combined_data_01deg, 
            aes(x = lon_std, y = lat_std, fill = predicted_fishing_hours)) +
  scale_fill_viridis(
    option = "inferno",
    direction = -1,
    trans = "log1p",
    name = "AIS fishing effort (hours; 2017-2020)", 
    breaks = c(0, 1, 10, 100, 1000, 10000, 100000, 1000000),
    labels = scales::comma,
    guide = guide_colorbar(barwidth = 20, barheight = 0.5, title.position = "top", title.hjust = 0.5)
  ) +
  coord_fixed(1.3) +
  theme_minimal() +
  labs(title = "Global fishing hours (0.1 degree resolution)",
       subtitle = "Based on AIS data and random forest predictions from SAR data",
       x = "Longitude", y = "Latitude") +
  theme(
    legend.position = "bottom",
    legend.direction = "horizontal",
    legend.box = "vertical",
    legend.margin = ggplot2::margin(t = 20, r = 0, b = 0, l = 0),
    legend.title = element_text(margin = ggplot2::margin(b = 10))
  )

print(predicted_plot)

# Save the plot
ggsave(here::here("R", "Outputs", "predicted_plot_SAR_only.png"), 
       plot = predicted_plot,
       width = 12, 
       height = 8, 
       dpi = 300)